Interactive NLP Chatbot for Accelerating Internal Operations

For this task, the NIX team has chosen the Tensorflow Hub implementation of the widely used BERT natural language model and implemented a classifier. It was a best-fitting solution- robust framework, strong pre-trained language models, and ability to easily apply and modify the Keras interface for Tensorflow.

It also provides the ability to join the pre-processing model with the primary model (so that we can directly send text requests to the Tensorflow Serving Server without any intermediate steps) and allows for an easy way to deploy trained models.

Our practice has shown that a pre-trained BERT is excellent at finding the correct answers to the questions contained in the documents. The difficulty lies directly in processing the documents themselves. Documents are HTML pages with a complex nested structure, which makes it challenging to find the correct answer to a question.

For this purpose, we developed a separate module that parses the documents and converts them into a flat format while preserving their contextual structure and richness. We also used the BERT model to search for the most relevant documents.